15-740/18-740 Computer Architecture Lecture 28: SIMD and GPUs Prof. Onur Mutlu Carnegie Mellon University

Announcements Project Poster Session Project Report Due December 10 NSH Atrium 2:30-6:30pm Project Report Due December 12 The report should be like a good conference paper Focus on Projects All group members should contribute Use the milestone feedback from the TAs

Final Project Report and Logistics Follow the guidelines in project handout for report We will provide the Latex format Good papers should be similar to the best conference papers you have been reading throughout the semester Submit all code, documentation, supporting documents and data Provide instructions as to how to compile and use your code in a README file This will determine part of your grade We will provide the directory to upload This is the single most important part of the project

Best Projects Best projects will be encouraged for a top conference submission Talk with me if you are interested in this Examples from past: Yoongu Kim, Dongsu Han, Onur Mutlu, and Mor Harchol-Balter, "ATLAS: A Scalable and High-Performance Scheduling Algorithm for Multiple Memory Controllers,” HPCA 2010 Best Paper Session George Nychis, Chris Fallin, Thomas Moscibroda, and Onur Mutlu, "Next Generation On-Chip Networks: What Kind of Congestion Control Do We Need?,” HotNets 2010. Yoongu Kim, Michael Papamichael, Onur Mutlu, and Mor Harchol-Balter, "Thread Cluster Memory Scheduling: Exploiting Differences in Memory Access Behavior,” MICRO 2010. (IEEE Micro Top Picks 2010)

Please Fill Out Course Evaluations Please fill them out online until December 14, 4pm Very important for feedback, course development/improvement, administration I read each of these carefully to improve the future course contents, logistics, etc. http://cmu.onlinecourseevaluations.com

TA Evaluations Please fill them out online until December 10, 5pm Vivek Seshadri: http://www.surveymonkey.com/s/PRW7DDJ Lavanya Subramanian: http://www.surveymonkey.com/s/PRTCNBD Evangelos Vlachos: http://www.surveymonkey.com/s/XZ88LVF

Last Time VLIW EPIC: Explicitly Parallel Instruction Computing Concepts and Philosophy Encoding and NOPs Static Scheduling Concepts Trace Scheduling Superblock Scheduling Hyperblock Scheduling EPIC: Explicitly Parallel Instruction Computing IA-64 Static store-load scheduling

Today Data Parallel (SIMD) Execution Model GPU Basics GPU Programming  18-742

Readings: SIMD and GPUs Required Lindholm et al., “NVIDIA Tesla: A Unified Graphics and Computing Architecture,” IEEE Micro 2008. Russell, “The CRAY-1 computer system,” CACM 1978. Recommended Fung et al., “Dynamic Warp Formation and Scheduling for Efficient GPU Control Flow,” MICRO 2007. Luk et al., “Qilin: Exploiting Parallelism on Heterogeneous Multiprocessors with Adaptive Mapping,” MICRO 2009.

Data Parallelism Concurrency arises from performing the same operations on different pieces of data Single instruction multiple data (SIMD) E.g., dot product of two vectors Contrast with thread (“control”) parallelism Concurrency arises from executing different threads of control in parallel Contrast with data flow Concurrency arises from executing different operations in parallel (in a data driven manner) SIMD exploits instruction-level parallelism Multiple instructions concurrent: instructions happen to be the same

SIMD Processing Single instruction operates on multiple data elements In time or in space Multiple processing elements Time-space duality Array processor: Instruction operates on multiple data elements at the same time Vector processor: Instruction operates on multiple data elements in consecutive time steps

SIMD Processing Single instruction operates on multiple data elements In time or in space Multiple processing elements Time-space duality Array processor: Instruction operates on multiple data elements at the same time Vector processor: Instruction operates on multiple data elements in consecutive time steps

Array vs. Vector Processors ARRAY PROCESSOR VECTOR PROCESSOR Instruction Stream Same op @ same time Different ops @ time LD VR  A[3:0] ADD VR  VR, 1 MUL VR  VR, 2 ST A[3:0]  VR LD0 LD1 LD2 LD3 LD0 AD0 AD1 AD2 AD3 LD1 AD0 MU0 MU1 MU2 MU3 LD2 AD1 MU0 ST0 ST1 ST2 ST3 LD3 AD2 MU1 ST0 AD3 MU2 ST1 Different ops @ same space MU3 ST2 Time Same op @ space ST3 Space Space

SIMD Array Processing vs. VLIW

SIMD Array Processing vs. VLIW Array processor

Vector Processors A vector is a one-dimensional array of numbers Many scientific/commercial programs use vectors for (i = 0; i<=49; i++) C[i] = (A[i] + B[i]) / 2 A vector processor is one whose instructions operate on vectors rather than scalar (single data) values Basic requirements Need to load/store vectors  vector registers (contain vectors) Need to operate on vectors of different lengths  vector length register (VLEN) Elements of a vector might be stored apart from each other in memory  vector stride register (VSTR) Stride: distance between two elements of a vector

Vector Processors (II) A vector instruction performs an operation on each element in consecutive cycles Vector functional units are pipelined Each pipeline stage operates on a different data element Vector instructions allow deeper pipelines No intra-vector dependencies  no hardware interlocking within a vector No control flow within a vector Known stride allows prefetching of vectors into memory

Vector Processor Advantages + No dependencies within a vector Pipelining, parallelization work well Can have very deep pipelines, no dependencies! + Each instruction generates a lot of work Reduces instruction fetch bandwidth + Highly regular memory access pattern Interleaving multiple banks for higher memory bandwidth Prefetching + No need to explicitly code loops Fewer branches in the instruction sequence

Vector ISA Advantages Compact encoding one short instruction encodes N operations Expressive, tells hardware that these N operations: are independent use the same functional unit access disjoint registers access registers in same pattern as previous instructions access a contiguous block of memory (unit-stride load/store) access memory in a known pattern (strided load/store) Scalable can run the same code in parallel pipelines (lanes)

Vector Processor Disadvantages -- Works (only) if parallelism is regular (data/SIMD parallelism) ++ Vector operations -- Very inefficient if parallelism is irregular -- How about searching for a key in a linked list?

Vector Processor Limitations -- Memory (bandwidth) can easily become a bottleneck, especially if 1. compute/memory operation balance is not maintained 2. data is not mapped appropriately to memory banks

Vector Functional Units Use deep pipeline (=> fast clock) to execute element operations Simplifies control of deep pipeline because elements in vector are independent (=> no hazards!) V1 V2 V3 Six stage multiply pipeline V3 <- v1 * v2 Slide credit: Krste Asanovic

Vector Instruction Execution ADDV C,A,B C[1] C[2] C[0] A[3] B[3] A[4] B[4] A[5] B[5] A[6] B[6] Execution using one pipelined functional unit C[4] C[8] C[0] A[12] B[12] A[16] B[16] A[20] B[20] A[24] B[24] C[5] C[9] C[1] A[13] B[13] A[17] B[17] A[21] B[21] A[25] B[25] C[6] C[10] C[2] A[14] B[14] A[18] B[18] A[22] B[22] A[26] B[26] C[7] C[11] C[3] A[15] B[15] A[19] B[19] A[23] B[23] A[27] B[27] Execution using four pipelined functional units Slide credit: Krste Asanovic

Vector Memory System 1 2 3 4 5 6 7 8 9 A B C D E F + Cray-1, 16 banks, 4 cycle bank busy time, 12 cycle latency Bank busy time: Cycles between accesses to same bank 1 2 3 4 5 6 7 8 9 A B C D E F + Base Stride Vector Registers Memory Banks Address Generator Slide credit: Krste Asanovic

Vector Unit Structure Functional Unit Vector Registers Lane Elements 0, 4, 8, … Elements 1, 5, 9, … Elements 2, 6, 10, … Elements 3, 7, 11, … Memory Subsystem Slide credit: Krste Asanovic

Vector Instruction Level Parallelism Can overlap execution of multiple vector instructions example machine has 32 elements per vector register and 8 lanes Complete 24 operations/cycle while issuing 1 short instruction/cycle Load Unit Multiply Unit Add Unit load mul add time load mul add Instruction issue Slide credit: Krste Asanovic

Vector Registers Each vector data register holds N M-bit values Vector control registers: VLEN, VSTR, VMASK Vector Mask Register (VMASK) Indicates which elements of vector to operate on Set by vector test instructions e.g., VMASK[i] = (Vk[i] == 0) Maximum VLEN can be N Maximum number of elements stored in a vector register M-bit wide M-bit wide V0,0 V1,0 V0,1 V1,1 V0,N-1 V1,N-1

Vector Machine Organization (CRAY-1) Russell, “The CRAY-1 computer system,” CACM 1978. Scalar and vector modes 8 64-element vector registers 64 bits per element 16 memory banks 8 64-bit scalar registers 8 24-bit address registers

Memory Banking in CRAY-1 Bank 1 Bank 2 Bank 15 MDR MAR MDR MAR MDR MAR MDR MAR Data bus Address bus CPU Slide credit: Derek Chiou

Scalar Code Example For I = 1 to 50 Scalar code C[i] = (A[i] + B[i]) / 2 Scalar code MOVI R0 = 50 1 MOVA R1 = A 1 MOVA R2 = B 1 MOVA R3 = C 1 X: LD R4 = MEM[R1++] 11 ;autoincrement addressing LD R5 = MEM[R2++] 11 ADD R6 = R4 + R5 4 SHFR R7 = R6 >> 1 1 ST MEM[R3++] = R7 11 DECBNZ --R0, X 2 ;decrement and branch if NZ 304 dynamic instructions

Scalar Code Execution Time Scalar execution time on an in-order processor with 1 bank First two loads in the loop cannot be pipelined 2*11 cycles 4 + 50*40 = 2004 cycles Scalar execution time on an in-order processor with 16 banks (word-interleaved) First two loads in the loop can be pipelined 4 + 50*30 = 1504 cycles Why 16 banks? 11 cycle memory access latency Having 16 (>11) banks ensures there are enough banks to overlap enough memory operations to cover memory latency

Vectorizable Loops A loop is vectorizable if each iteration is independent of any other For I = 0 to 49 C[i] = (A[i] + B[i]) / 2 Vectorized loop: MOVI VLEN = 50 1 MOVI VSTR = 1 1 VLD V0 = A 11 + VLN - 1 VLD V1 = B 11 + VLN – 1 VADD V2 = V0 + V1 4 + VLN - 1 VSHFR V3 = V2 >> 1 1 + VLN - 1 VST C = V3 11 + VLN – 1 7 dynamic instructions

Vector Code Performance No chaining i.e., output of a vector functional unit cannot be used as the input of another (i.e., no vector data forwarding) 16 memory banks (word-interleaved) 285 cycles

Vector Code Performance - Chaining Vector chaining: Data forwarding from one vector functional unit to another 182 cycles Each memory bank has a single port (memory bandwidth bottleneck) These two VLDs cannot be pipelined. WHY? VLD and VST cannot be pipelined. WHY?

Vector Chaining LV v1 MULV v3,v1,v2 ADDV v5, v3, v4 Memory V1 Load Unit Mult. V2 V3 Chain Add V4 V5 Chain LV v1 MULV v3,v1,v2 ADDV v5, v3, v4 Slide credit: Krste Asanovic

Vector Code Performance – Multiple Memory Ports Chaining and 2 load ports, 1 store port in each bank 79 cycles

Questions (I) What if # data elements > # elements in a vector register? Need to break loops so that each iteration operates on # elements in a vector register E.g., 527 data elements, 64-element VREGs 8 iterations where VLEN = 64 1 iteration where VLEN = 15 (need to change value of VLEN) Called vector stripmining What if vector data is not stored in a strided fashion in memory? (irregular memory access to a vector) Use indirection to combine elements into vector registers Called scatter/gather operations

Scatter/Gather Operations Want to vectorize loops with indirect accesses: for (i=0; i<N; i++) A[i] = B[i] + C[D[i]] Indexed load instruction (Gather) LV vD, rD # Load indices in D vector LVI vC, rC, vD # Load indirect from rC base LV vB, rB # Load B vector ADDV.D vA,vB,vC # Do add SV vA, rA # Store result

Scatter/Gather Operations Scatter/Gather operations often implemented in hardware to handle sparse matrices Vector loads and stores use an index vector which is added to the base register to generate the addresses Index Vector Data Vector Equivalent 1 3.14 3.14 3 6.5 0.0 7 71.2 6.5 8 2.71 0.0 0.0 71.2 2.7

Conditional Operations in a Loop What if some operations should not be executed on a vector (based on a dynamically-determined condition)? loop: if a[i] then b[i]=a[i]*b[i] goto loop Idea: Masked operations VMASK register is a bit mask determining which data element should not be acted upon VLD V0 = A VLD V1 = B VMASK = (V0 != 0) VMUL V1 = V0 * V1 VST B = V1 Does this look familiar? This is essentially predicated execution.

Another Example with Masking for (i = 0; i < 64; ++i) if (a[i] >= b[i]) then c[i] = a[i] else c[i] = b[i] Steps to execute loop 1. Compare A, B to get VMASK 2. Selective store of A,VMASK into C 3. Complement VMASK 4. Selective store of B, VMASK into C A B VMASK 1 2 0 2 2 1 3 2 1 4 10 0 -5 -4 0 0 -3 1 6 5 1 -7 -8 1

Masked Vector Instructions B[3] A[4] B[4] A[5] B[5] A[6] B[6] M[3]=0 M[4]=1 M[5]=1 M[6]=0 M[2]=0 M[1]=1 M[0]=0 Write data port Write Enable A[7] B[7] M[7]=1 Simple Implementation execute all N operations, turn off result writeback according to mask C[4] C[5] C[1] Write data port A[7] B[7] M[3]=0 M[4]=1 M[5]=1 M[6]=0 M[2]=0 M[1]=1 M[0]=0 M[7]=1 Density-Time Implementation scan mask vector and only execute elements with non-zero masks Slide credit: Krste Asanovic

Compress and Expand Operations Compress packs non-masked elements from one vector register contiguously at start of destination vector register population count of mask vector gives packed vector length Expand performs inverse operation Used for density-time conditionals and also for general selection operations M[3]=0 M[4]=1 M[5]=1 M[6]=0 M[2]=0 M[1]=1 M[0]=0 M[7]=1 A[3] A[4] A[5] A[6] A[7] A[0] A[1] A[2] M[3]=0 M[4]=1 M[5]=1 M[6]=0 M[2]=0 M[1]=1 M[0]=0 M[7]=1 B[3] A[4] A[5] B[6] A[7] B[0] A[1] B[2] Expand Compress Slide credit: Krste Asanovic

Reduction Operations Problem: Loop-carried dependence on reduction variables sum = 0; for (i=0; i<N; i++) sum += A[i]; # Loop-carried dependence on sum Solution: Re-associate operations if possible, use binary tree to perform reduction # Rearrange as: sum[0:VL-1] = 0 # Vector of VL partial sums for(i=0; i<N; i+=VL) # Stripmine VL-sized chunks sum[0:VL-1] += A[i:i+VL-1]; # Vector sum # Now have VL partial sums in one vector register do { VL = VL/2; # Halve vector length sum[0:VL-1] += sum[VL:2*VL-1] # Halve no. of partials } while (VL>1) Slide credit: Krste Asanovic

Automatic Code Vectorization for (i=0; i < N; i++) C[i] = A[i] + B[i]; Vector Instruction load add store Iter. 1 Iter. 2 Vectorized Code Time load add store Iter. 1 Iter. 2 Scalar Sequential Code Vectorization is a compile-time reordering of operation sequencing  requires extensive loop dependence analysis Slide credit: Krste Asanovic

Vector Processing Summary Vector machines good at exploiting regular data-level parallelism Same operation performed on many data elements Improve performance, simplify design (no intra-vector dependencies) Performance improvement limited by vectorizability of code Scalar operations limit vector machine performance Amdahl’s Law CRAY-1 was the fastest SCALAR machine at its time! Many existing ISAs include (vector-like) SIMD operations Intel MMX/SSEn, PowerPC AltiVec, ARM Advanced SIMD

Intel Pentium MMX Operations Idea: One instruction operates on multiple data elements simultaneously Ala array processing (yet much more limited) Designed with multimedia (graphics) operations in mind No VLEN register Opcode determines data type: 8 8-bit bytes 4 16-bit words 2 32-bit doublewords 1 64-bit quadword Stride always equal to 1. Peleg and Weiser, “MMX Technology Extension to the Intel Architecture,” IEEE Micro, 1996.

MMX Example: Image Overlaying (I)

MMX Example: Image Overlaying (II)

Graphics Processing Units

High-Level View of a GPU

Concept of “Thread Warps” Warp: A set of threads that execute the same instruction (on different data elements) All threads run the same kernel Warp: The threads that run lengthwise in a woven fabric … Thread Warp 3 Thread Warp 8 Common PC Thread Warp In SIMD, you need to specify the data array + an instruction (on which to operate the data on) + THE INSTRUCTION WIDTH. Eg: You might want to add 2 integer arrays of length 16, then a SIMD instruction would look like (the instruction has been cooked-up by me for demo) add.16 arr1 arr2 However, SIMT doesn't bother about the instruction width. So, essentially, you could write the above example as: arr1[i] + arr2[i] and then launch as many threads as the length of the array, as you want. Note that, if the array size was, let us say, 32, then SIMD EXPECTS you to explicitly call two such 'add.16' instructions! Whereas, this is not the case with SIMT. Scalar Scalar Scalar Scalar Thread Warp 7 Thread Thread Thread Thread W X Y Z SIMD Pipeline

Latency Hiding with “Thread Warps” Warp: A set of threads that execute the same instruction (on different data elements) Fine-grained multithreading One instruction per thread in pipeline at a time (No branch prediction) Interleave warp execution to hide latencies Register values of all threads stay in register file No OS context switching Memory latency hiding Graphics has millions of pixels Decode R F A L U D-Cache Thread Warp 6 Thread Warp 1 Thread Warp 2 Data All Hit? Miss? Warps accessing memory hierarchy Thread Warp 3 Thread Warp 8 Writeback Warps available for scheduling Thread Warp 7 I-Fetch SIMD Pipeline With a large number of shader threads multiplexed on the same execution re- sources, our architecture employs fine-grained multithreading where individual threads are interleaved by the fetch unit to proactively hide the potential latency of stalls before they occur. As illustrated by Figure, warps are issued fairly in a round-robin queue. When a thread is blocked by a memory request, shader core simply removes that thread’s warp from the pool of “ready” warps and thereby allows other threads to proceed while the memory system processes its request. With a large number of threads (1024 per shader core) interleaved on the same pipeline, FGMT effectively hides the latency of most memory operations since the pipeline is occupied with instructions from other threads while memory operations complete. also hides the pipeline latency so that data bypassing logic can potentially be omitted to save area with minimal impact on performance. simplify the dependency check logic design by restricting each thread to have at most one instruction running in the pipeline at any time. Slide credit: Tor Aamodt

Warp-based SIMD vs. Traditional SIMD Traditional SIMD contains a single thread Lock step Programming model is SIMD (no threads)  SW needs to know vector length ISA contains vector/SIMD instructions Warp-based SIMD consists of multiple scalar threads executing in a SIMD manner (i.e., same instruction executed by all threads) Does not have to be lock step Each thread can be treated individually (i.e., placed in a different warp)  programming model not SIMD SW does not need to know vector length Enables memory and branch latency tolerance ISA is scalar  vector instructions formed dynamically Essentially, it is MIMD/SPMD programming model implemented on SIMD hardware

Branch Divergence Problem in Warp-based SIMD SPMD Execution on SIMD Hardware NVIDIA calls this “Single Instruction, Multiple Thread” (“SIMT”) execution B C D E F A G Thread Warp Common PC Thread 1 Thread 2 Thread 3 Thread 4 Slide credit: Tor Aamodt

Control Flow Problem in GPUs/SIMD GPU uses SIMD pipeline to save area on control logic. Group scalar threads into warps Branch divergence occurs when threads inside warps branch to different execution paths. Branch Path A Path B Branch Path A Path B Slide credit: Tor Aamodt

Branch Divergence Handling (I) Reconv. PC Next PC Active Mask Stack B C D E F A G A/1111 E D 0110 C 1001 TOS - 1111 - E 1111 TOS E D 0110 1001 TOS - 1111 - A 1111 TOS - B 1111 TOS E D 0110 TOS - 1111 - G 1111 TOS B/1111 C/1001 D/0110 Thread Warp Common PC Thread 2 3 4 1 E/1111 G/1111 A B C D E G A Time Slide credit: Tor Aamodt

Branch Divergence Handling (II) if (some condition) { B; } else { C; } D; One per warp Control Flow Stack Next PC Recv PC Amask TOS A -- 1111 D -- 1111 B D 1110 A C D 0001 D Execution Sequence 1 A 1 C 1 B 1 D B C D Time Slide credit: Tor Aamodt

Dynamic Warp Formation Idea: Dynamically merge threads executing the same instruction (after branch divergence) Form new warp at divergence Enough threads branching to each path to create full new warps

Dynamic Warp Formation/Merging Idea: Dynamically merge threads executing the same instruction (after branch divergence) Fung et al., “Dynamic Warp Formation and Scheduling for Efficient GPU Control Flow,” MICRO 2007. Branch Path A Path B Branch Path A

Dynamic Warp Formation Example Execution of Warp x at Basic Block A Execution of Warp y Legend A B x/1110 y/0011 C x/1000 D x/0110 F x/0001 y/0010 y/0001 y/1100 A new warp created from scalar threads of both Warp x and y executing at Basic Block D D E x/1110 y/0011 G x/1111 y/1111 A A B B C C D D E E F F G G A A Baseline Time Dynamic Warp Formation A A B B C D E E F G G A A Time Slide credit: Tor Aamodt

How to Fill Holes in Warps?

Memory Access within A Warp “To improve memory bandwidth and reduce overhead, the local and global load/ store instructions coalesce individual parallel thread accesses from the same warp into fewer memory block accesses.” Highest efficiency achieved if individual threads within a warp access consecutive locations in memory  same row If threads within a warp conflict with each other, SIMD efficiency degrades significantly  similar to traditional SIMD machines

What About Memory Divergence? Modern GPUs have caches Ideally: Want all threads in the warp to hit (without conflicting with each other) Problem: One thread in a warp can stall the entire warp if it misses in the cache. Dynamic warp formation can cause bank conflicts between threads within a warp (if the warp is not formed in a bank-aware manner) Need techniques to Tolerate memory divergence Integrate solutions to branch and memory divergence

NVIDIA GeForce GTX 285 NVIDIA-speak: 240 stream processors “SIMT execution” Generic speak: 30 cores 8 SIMD functional units per core Slide credit: Kayvon Fatahalian

NVIDIA GeForce GTX 285 “core” … 64 KB of storage for fragment contexts (registers) 30 * 32 * 32 = 30 * 1024 = 30K fragments 64KB register file = 16 32-bit registers per thread = 64B (1/32 that of LRB) 16KB of shared scratch 80KB / core available to software = SIMD functional unit, control shared across 8 units = instruction stream decode = multiply-add = execution context storage = multiply Slide credit: Kayvon Fatahalian

NVIDIA GeForce GTX 285 “core” … 64 KB of storage for fragment contexts (registers) To get maximal latency hiding: Run 1/32 of the time 16 words per thread = 64B Groups of 32 [fragments/vertices/threads/etc.] share instruction stream (each group is a Warp) Up to 32 warps are simultaneously interleaved Up to 1024 thread contexts can be stored Slide credit: Kayvon Fatahalian

There are 30 of these things on the GTX 285: 30,720 threads NVIDIA GeForce GTX 285 Tex Tex … … … … Tex Tex … … Tex Tex … … Tex Tex … If you’re running a CUDA program, and your not launching 30K threads, you are certainly not getting full latency hiding, and you might not be using the GPU well … Tex Tex … … There are 30 of these things on the GTX 285: 30,720 threads Slide credit: Kayvon Fatahalian

A More Detailed View Lindholm et al., “NVIDIA Tesla: A Unified Graphics and Computing Architecture,” IEEE Micro 2008.

NVIDIA GeForce GTX 285 Generic speak: 30 processing cores 8 SIMD functional units per core Best case: 240 mul-adds + 240 muls per clock Tons of latency hiding At the cost of tiny amounts of per fragment storage

Food for Thought